Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759

Interactive Deep Learning System for

Automated Car Damage Detection: Multi-Model
Evaluation and Interactive Web Deployment
Sai Madhu1; Bharathi Maddikatla2; Ranjitha Padakanti3;
Vineel Sai Kumar Rampally4; Shirish Kumar Gonala5
1
Research Intern, Innodatatics, Hyderabad, India.
2
Associate Data Scientist, Innodatatics, Hyderabad, India.
3
Senior Data Scientist, Innodatatics, Hyderabad, India.
4
Salesforce Solution Architect, Fresenius Medical Care, North America
5
Founder and CEO, Innodatatics, Hyderabad, India

ORCID- 0009-0009-3701-8929

Publication Date: 2025/05/08

Abstract: This project presents an automated framework for vehicle damage evaluation employing deep learning
methodologies, designed to optimize assessment procedures within automotive service environments. By implementing the
YOLOv9 computational vision architecture, the system enables rapid identification of vehicular damage components
through advanced pattern recognition, reducing reliance on labor-intensive manual inspections. The model underwent
training on an extensive curated dataset comprising 8,450 annotated images capturing diverse damage morphologies across
multiple vehicle perspectives, including frontal collisions, lateral impacts, and rear-end accidents. The framework integrates
physics-informed augmentation strategies to enhance environmental adaptability, particularly addressing challenges posed
by variable lighting conditions and reflective surfaces. A modular processing pipeline facilitates scalable deployment through
quantization techniques optimized for edge computing devices, demonstrating practical applicability in service center
operations. The system incorporates a web-based interface enabling real-time damage visualization and automated report
generation, significantly streamlining technician workflows. Experimental results indicate substantial improvements in
inspection efficiency, with the YOLOv9 architecture achieving 87% mean average precision ([email protected]) while maintaining
computational efficiency. Quantized model variants exhibited a 68% reduction in memory footprint with minimal accuracy
degradation. Field validations conducted across multiple service centers confirmed the system's operational effectiveness,
highlighting strong correlations between model complexity, training duration, and real-time detection capabilities. This
research establishes foundational insights for future advancements in 3D damage reconstruction and adaptive learning
systems within automotive diagnostics.

Keywords: Computational Damage Assessment, YOLOv9 Architecture, Automotive Computer Vision, Edge AI Optimization, Service
Process Automation, Neural Network Quantization.

How to Cite: Sai Madhu; Bharathi Maddikatla; Ranjitha Padakanti; Vineel Sai Kumar Rampally; Shirish Kumar Gonala (2025).
Interactive Deep Learning System for Automated Car Damage Detection: Multi-Model Evaluation and Interactive
Web Deployment. International Journal of Innovative Science and Research Technology,
10(4), 2779-2798. https://doi.org/10.38124/ijisrt/25apr1759

I. INTRODUCTION resulting in delayed repair timelines and insurance

processing.
The automotive service sector is undergoing rapid
transformation through the integration of artificial This study addresses these operational challenges
intelligence and machine vision technologies. [2]. through the development of an automated deep learning
Conventional damage assessment methodologies remain framework employing YOLO object detection architecture.
constrained by subjective human evaluation, inconsistent [1, 6]. The system enables real-time damage localisation
documentation practices, and temporal inefficiencies, often across 22 distinct structural and component failure categories,
trained on 8,450 annotated vehicle images encompassing

IJISRT25APR1759 www.ijisrt.com 2779

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759
diverse collision scenarios [1, 6]. Integration of physics-based business objective alignment, data acquisition, model
data augmentation enhances model generalisation under development, quantitative evaluation, deployment
variable environmental conditions [1], while CRISP-ML(Q) optimisation, and continuous monitoring [1]. This structured
methodology ensures rigorous quality assurance throughout approach facilitates traceability and quality control across all
the machine learning lifecycle. system components, from initial data collection to production
deployment.
As illustrated in [Fig.1], the CRISP-ML(Q) framework
governs project execution through six iterative phases:

Fig 1 CRISP-ML(Q) Methodology for Proposed Damage Detection System

IJISRT25APR1759 www.ijisrt.com 2780

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759
II. METHODS AND TECHNIQUES

Fig 2 Architecture Diagram: Showcasing the Components and Flow of Data

 Business Understanding and localization of car damages to assist service centers in

In modern automotive service centers, damage automating their inspection workflows.
assessment is often performed manually, leading to
inefficiencies, subjective interpretations, and delayed service  Success Criteria:
delivery. The lack of automation in the initial inspection
process directly affects repair timelines, cost estimation  Achieve a damage detection accuracy of over 90% on
accuracy and customer satisfaction. validation datasets.
 Reduce manual inspection time by at least 50%.
To address these operational challenges, this project  Enable automated report generation with localised
proposes an AI-powered vehicle damage detection system damage annotations.
using the YOLOv9 object detection model. The system is  Ensure the model runs efficiently on edge devices with a
designed to analyze images of vehicles at the time of intake quantised implementation.
and automatically detect visible external damages such as
dents, scratches and cracks. By aligning these criteria with service centre workflows,
the solution supports both technical innovation and
 Objective: operational improvement.
The primary objective is to develop a robust deep
learning framework that enables real-time, accurate detection

IJISRT25APR1759 www.ijisrt.com 2781

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759
III. DATA COLLECTION detection and ensures comprehensive coverage of various
vehicle models, damage types, and real-world conditions [3].
To develop a robust and generalisable vehicle damage
detection model, we curated a comprehensive dataset by  Data Collection Process
aggregating images from multiple sources: publicly available
internet datasets, proprietary datasets shared by clients, and  Internet Sources: 4,900 images from public repositories
real-world images collected from various car service centres. (e.g., Roboflow Universe, Kaggle), covering 23
As illustrated in [Fig. 2], the goal was to ensure diversity in standardized damage categories.
damage types, vehicle models, angles [1], lighting conditions  Client Data: 3,550 proprietary images from insurance
[6], and image resolutions, reflecting realistic service centre claims, annotated for part-specific damages (e.g., bumper,
environments. windscreen).
 Service Center Images: 1,350 high-resolution images
A. Data Sources captured during routine inspections, enhancing dataset
The dataset for vehicle damage detection was diversity under varied lighting and environmental
constructed from three primary sources: (1) publicly available conditions [1].
internet datasets, (2) proprietary client data, and (3) real-
world images collected directly from service centres. Videos from all sources were converted to frames at 5
Illustrated in [Fig. 2], this multi-source strategy addresses the FPS using FFmpeg, adding 2,000 video-derived images to the
challenge of limited public datasets for car centre damage dataset.

Table 1 Overview of Data Sources Used for Vehicle Damage Image Collection

Following comprehensive data cleaning and paired with detailed annotations, and summary tables present
augmentation steps as depicted in [Fig.2], the finalised dataset the class distributions and image characteristics, supporting
consists of 8,450 labelled images spanning 22 categories of transparent reporting and reproducibility in deep learning
vehicle damage. This dataset reflects a broad spectrum of research [11]. This thorough documentation ensures that the
real-world damage scenarios and vehicle types, which is dataset can serve as a robust foundation for both model
essential for training models that generalise effectively to development and future benchmarking efforts within the field
practical automotive service environments. Each image is of automated vehicle damage detection.

Table 2 Label-wise Count for Vehicle Damage Detection

Label Count
dent 145
Glass_Break 86
quaterpanel-dent 401
Taillight-Damage 236
doorouter-dent 613
Rear-windscreen-Damage 258
rear-bumper-dent 560
Sidemirror-Damage 210
front-bumper-dent 1044

IJISRT25APR1759 www.ijisrt.com 2782

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759
Label Count
dent 145
Glass_Break 86
bonnet-dent 577
fender-dent 495
Scratch 153
pillar-dent 33
Signlight-Damage 63
RunningBoard-Dent 174
roof-dent 194
scratch 181
Dent 250
glass 9
Headlight-Damage 344
medium-Bodypanel-Dent 3
Front-Windscreen-Damage 167

 List of Car Damages systems (headlights, taillights, signlights), and mirrors.

Component failures are essential to vehicle safety and thus
 Structural Damages: are prioritised in automated assessment workflows. The
Structural damages encompass issues such as dents, dataset incorporates images captured under diverse lighting
cracks, and deformations affecting panels like bumpers, conditions and from multiple perspectives, which enhances
doors, fenders, and roofs. These types of damage typically the model’s ability to generalise and perform reliably in real-
result from collisions or impacts and are among the most world service scenarios [6].
frequently reported in insurance claims. The dataset used in
this study includes comprehensive annotations and a wide  Surface Imperfections:
variety of images, as detailed in [Table 2] and illustrated by Surface imperfections, such as scratches and paint
sample images, ensuring that different severities, types, and damage, are frequent but often challenging to detect due to
locations of structural damage are well represented for robust their subtle appearance and sensitivity to lighting variations.
model training and evaluation [1, 3]. To address this, the dataset features images with a range of
backgrounds, climates, and lighting conditions, thereby
 Component Failures: improving the model’s robustness and ability to identify these
This category includes damages to critical vehicle less conspicuous forms of damage [1, 6].
components such as glass (windshield, rear window), lighting

Table 3 Data Split of the Dataset

Split Image Count
Train 3808
Vali 441
Test 18

B. Data Preprocessing computational efficiency and preservation of fine damage

A robust preprocessing pipeline is essential for details. This input size was chosen to comply with YOLO
optimising deep learning models in vehicle damage detection. architecture requirements and to maintain consistency across
Our approach integrates best practices from recent studies, the dataset, which includes 8,450 images spanning 22 damage
including resolution, colour normalisation, and targeted classes [12]. The 640×640 resolution has been shown to offer
augmentation strategies tailored for automotive imagery [6]. an optimal trade-off between detection accuracy and
As shown in [Table.2], the dataset comprises images from processing speed for real-time automotive damage
multiple sources, such as service centres, which contributed assessment [1, 8]. As illustrated in [Fig.2], this standardised
3,100 high-resolution images under diverse lighting and input size feeds into the CSPDarknet backbone, facilitating
environmental conditions, thereby improving dataset consistent feature extraction throughout the model pipeline.
diversity and model robustness, data augmentation is used to Bilinear interpolation was preferred over nearest-neighbour
artificially expand and adapt our datasets [3]. methods, as it better preserves texture details necessary for
distinguishing between similar damage types, such as
 Image Resizing scratches and paint chips.
All images were resized to 640×640 pixels using
bilinear interpolation, ensuring a balance between

IJISRT25APR1759 www.ijisrt.com 2783

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759

Fig 3 Appearance of Damages in Different Lighting Variations

 Colour Space Transformation enhances the model's ability to generalize across different
To reduce the impact of lighting variability, pixel values lighting conditions ,As illustrated in [Fig.3], which is
were normalized by dividing by 255, bringing them into a particularly relevant for reflective automotive surfaces [6].
range. This normalization step stabilizes training and

IJISRT25APR1759 www.ijisrt.com 2784

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759

Fig 4 Data Augmentation of Damaged Cars

 Image Normalization the model focus on relevant features rather than variations
Further normalization was performed by standardizing caused by lighting or sensor differences, as recommended in
pixel intensity distributions using mean and standard prior vehicle damage detection research [6].
deviation calculated from the training set. This process helps

Table 4 Comparison of Normalization Techniques

Our experiments demonstrated that standard applied. This step enhances the visibility of subtle damages,
normalization outperformed other techniques such as Z-score such as fine scratches or minor dents, without introducing
normalization for automotive damage detection. This finding artifacts that could mislead the model [12].
aligns with [6], who observed that maintaining the positive
range of pixel values preserves important visual  Background Standardization
characteristics of damage patterns while still providing the Images often contain complex backgrounds that can
benefits of normalization. distract the model. A combination of semantic segmentation
and selective blurring was used to de-emphasize non-vehicle
 Noise Reduction and Image Enhancement regions, helping the model focus on relevant damage areas.
To improve clarity and reduce the influence of noise, This approach has been shown to reduce false positives and
especially in images captured under suboptimal conditions, improve detection rates in automotive datasets.
adaptive bilateral filtering and histogram equalization were

IJISRT25APR1759 www.ijisrt.com 2785

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759
Table 5 Impact of Different Background Standardization Techniques

As shown in [Table.4], selective background blurring  Rotation

yielded the best results, improving mAP by 2.8% while Random rotations up to ±20° were used to simulate
reducing false positives by 15%. This approach preserves various camera angles encountered in real-world inspections,
important contextual information while minimising the reducing the risk of angle-dependent misclassification [10].
influence of irrelevant background variations, allowing the
model to focus more effectively on actual vehicle damages  Flip
[4]. Horizontal flipping was applied with a 50% probability
to increase the diversity of damage orientations, especially for
 Data Augmentation symmetrical vehicle parts [7].
To improve model generalisation and address class
imbalance, several augmentation techniques were applied as  Shear
illustrated in [Fig. 5]: Shear transformations up to 15% were introduced to
mimic perspective distortions, further enhancing the model’s
robustness to different viewpoints.

Table 6 Data Augmentation Techniques and their impact on Model

As illustrated in [Table.6], shear augmentation sampling. This approach is essential for mitigating class
contributed to a 1.8% improvement in mAP, primarily by imbalance, which can otherwise bias model performance and
enhancing the model's robustness to perspective variations. reduce its ability to generalise to under-represented
This augmentation was particularly effective for improving categories. We adopted an 85:10:5 split ratio, allocating 85%
detection performance on damages captured at oblique of the data for training, 10% for validation during model
angles, which are common in real-world inspection scenarios development, and 5% for final testing and evaluation. As
[6]. shown in [Table.7], this strategy ensures that each subset
contains a proportional distribution of all damage classes,
 Data Splitting supporting robust model training and unbiased performance
To ensure balanced representation of all damage assessment. Stratified sampling has been widely
classes—including rare types such as "Medium-Bodypanel- recommended in automotive damage detection literature to
Dent" with as few as three instances—the dataset was divided maintain dataset integrity and support fair evaluation of deep
into training, validation, and testing sets using stratified learning models [12].

IJISRT25APR1759 www.ijisrt.com 2786

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759
Table 7 Dataset Split Distribution across Training, Validation, and Testing Sets

This approach ensures that rare damage types are proportionally represented in each subset, supporting reliable evaluation and
minimizing bias.

IV. MODEL BUILDING

This section details the implementation and performance of various YOLO (You Only Look Once) architectures for car damage
identification. We evaluated multiple YOLO variants with different configurations to identify and verify the optimal model for
classification and identification of vehicle damages across 22 damage classes.

Fig 5 YOLOV Architecture

IJISRT25APR1759 www.ijisrt.com 2787

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759
A. YOLOv8 precision). However, it struggled with subtle damage types
YOLOv8 represents a significant advancement in object such as scratches and minor glass cracks, achieving only 61%
detection architectures providing improved accuracy along and 58% precision, respectively.
with speed of solution compared to previous generations [2].
"YOLOv8 was chosen because of its real-time detection C. Advanced YOLO Variants
capabilities, which make it appropriate for real-world uses We also evaluated several newer and experimental
like automated inspections and processing insurance claims." YOLO variants to assess high level performance
improvements for car damage identification:
 The YOLOv8 Architecture Incorporates Several Key
Features that Enhance its Performance for Car Damage  YOLOv9
Detection: YOLOv9 demonstrated the most promising results
among advanced variants, with accuracy improving
 CSP (Cross-Stage Partial) backbone: Efficiently extracts significantly as training epochs increased. The 250-epoch
features while reducing computational requirements. YOLOv9s model achieved our highest overall accuracy of
 SPPF (Spatial Pyramid Pooling Fast): Improves 87% [email protected], suggesting architectural improvements
identification of damages of varying sizes [13]. provide meaningful benefits for car damage detection tasks.
 Decoupled head: Provides better localisation and
classification of damage types [13].  Key Innovations in YOLOv9 Include:

Our experimentations with YOLOv8 variants revealed  Swin Transformer backbone: Enhances feature extraction
promising results, with the nano (n) and small (s) versions with attention mechanisms [2].
demonstrating an excellent maintenance between accuracy  Gradient flow optimization: Improves training stability
and computational efficiency. The YOLOv8s model achieved and convergence [4].
83% [email protected] when trained for 50–100 epochs, benefiting  Dynamic label assignment: Better handles the variety of
from its larger parameter count (11.2M) while maintaining damage types and sizes [1].
reasonable inference speed (72 FPS on NVIDIA RTX
3080ti). As illustrated in [Fig. 5] YOLOv9's attention
mechanisms allow it to focus more effectively on damage
As shown in the architecture diagram [Fig.5], regions while suppressing background noise, which is
YOLOv8's feature pyramid network effectively captures particularly valuable for detecting subtle damages like
multi-scale features, which is critical for detecting damages scratches against complex vehicle surfaces.
ranging from small scratches to large dents. The model's
ability to process images at 640×640 resolution provides  YOLOv11 and YOLOv12
sufficient detail for accurate damage localization while
maintaining efficiency [8].  YOLOv11 and YOLOv12 Variants Showed Mixed Results:

Extended training of YOLOv8n (0–150 epochs)  YOLOv12m achieved competitive accuracy (78%
demonstrated significant improvement over the 50-epoch [email protected]) comparable to YOLOv8s but required
training, highlighting the importance of sufficient training significantly more parameters (25.6M vs. 11.2M).
iterations for complex damage classification tasks [Table.8].  Hybrid CNN-Transformer design: Improved multi-scale
where longer training periods consistently yielded better feature extraction but offered marginal performance gains
results. [2].

B. YOLOv7  Performance Analysis

YOLOv7 represents an earlier generation in the YOLO Across all tested models, several patterns emerged:
family but remains competitive for specific applications [4].
YOLOv7 "significantly outperformed its predecessors  Model size impact: Larger models (s/m variants)
regarding performance by implementing training techniques consistently outperformed nano counterparts, suggesting
and architectural enhancements." increased model capacity benefits complex damage
detection.
While we focused more extensively on newer YOLO  Training duration: Extended training periods (e.g.,
variants, limited testing with YOLOv7 showed promising YOLOv9s from 50 to 250 epochs) improved mAP by
results for larger, more visible damage types such as major 10%, highlighting the need for longer schedules due to
dents and broken components. The architecture's E-ELAN subtle visual differences between damage categories [3].
(Extended Efficient Layer Aggregation Network) backbone  Architectural advantages: YOLOv9's Swin Transformer
provides effective feature extraction capabilities, though it backbone outperformed earlier architectures, particularly
lacks refinements found in later YOLO versions [1]. in distinguishing similar damage types [11].
 Inference speed tradeoffs: Nano variants maintained >90
We found that YOLOv7 achieved 76% mAP when FPS on RTX 3080ti GPUs, while small variants operated
trained for 100 epochs, with strong performance on front at 45–72 FPS, enabling deployment flexibility [Table.8].
bumper damage (85% precision) and door dents (82%

IJISRT25APR1759 www.ijisrt.com 2788

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759
V. EVALUATION

This section analyzes the performance of YOLO-based architectures for automotive damage identification, focusing on
accuracy-efficiency trade-offs and real-time deployment considerations.

Fig 6 Recall-Confidence curves for each Damage Class

Fig 7 Precision-Recall curves for 22 Damage Classes

IJISRT25APR1759 www.ijisrt.com 2789

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759

Fig 8a Confusion Matrix for car Damage across 22 Damage Classes

Fig 8b Normalized Confusion Matrix for each True Class

IJISRT25APR1759 www.ijisrt.com 2790

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759
 Model Comparison
We evaluated YOLO variants using key metrics: mean Average Precision ([email protected]), inference speed, and computational
specifications [Table.3].

Table 8 Performance Comparison of YOLO Variants for Car Damage Detection

As shown in [Table.8], YOLOv9s trained for 250 and architectural innovations like Swin Transformers
epochs achieved the highest mAP (87.0%), demonstrating a contributed to this performance leap while maintaining real-
47% improvement over YOLOv8n. Extended training cycles time capabilities (53.2 FPS) [6].

Fig 9 Precision-Confidence curves for Different Damage Categories

IJISRT25APR1759 www.ijisrt.com 2791

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759
 Class-wise Performance Analysis  Low accuracy (<70%): Glass cracks (67.2%), Scratches
Analysis revealed significant variance across 22 damage (63.5%), Medium body panel dents (42.1%) [Fig.9].
classes:
Class imbalance in training data and visual similarity
 High accuracy (>85%): Front bumper dents (91.2%), between type of damages primarily drove these
Hood dents (89.7%), Rear bumper dents (87.3%) discrepancies, "parameter optimization and dataset expansion
 Moderate accuracy (70-85%): Door panel dents (82.1%), remain critical for improving rare damage recognition [3]."
Fender dents (78.9%), Headlight damage (75.4%)

Fig 10 F1-Confidence curves for each Damage Class

 Confusion Matrix Analysis [Fig.3] demonstrates YOLOv9s' robustness across

Confusion matrices identified key misclassification damage types, significantly in localizing bumper dents under
trends: varying lighting. The integration of transformer-based
attention improved boundary precision by 12% compared to
 Scratch-Dent Confusion: 23% mislabeling on door panels CNN-only architectures [4].
due to overlapping texture features
 Glass-Light Damage: 18% confusion from reflective This evaluation framework provides actionable insights
surface artifacts for selecting models based on operational requirements,
 Adjacent Component Errors: 15% merging of fender/door balancing identification accuracy with constraints in
dent predictions [Fig.8a, Fig.8b]. automotive inspection systems.
 These patterns align with, who emphasized that "subtle
geometric differences require enhanced spatial attention VI. DEPLOYMENT STRATEGY
mechanisms in feature extraction [1]."
Our deployment strategy focuses on creating an
 Optimal Model Selection accessible, scalable system that enables service centers to
Based on comprehensive testing: efficiently integrate car damage detection into their
workflows. We developed a web-based interface using
 YOLOv9s (250 epochs): Optimal choice for accuracy- Streamlit, allowing technicians to upload images, visualize
critical applications (87.0% mAP @53.2 FPS) detections, and generate standardized reports.
 YOLOv8n (150 epochs): Preferred for edge devices (81%
mAP @72 FPS) with minimal parameter overhead (3.2M)  Framework Selection
After evaluating multiple deployment frameworks, we
selected Streamlit for its balance of simplicity, flexibility, and

IJISRT25APR1759 www.ijisrt.com 2792

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759
interactive capabilities. As shown in [Fig.2], the deployment  Model Optimization for Deployment
architecture consists of: To ensure efficient deployment on edge devices with
varying computational capabilities, we implemented several
 Frontend Interface: Streamlit-based web application optimization techniques:
providing image upload, visualization, and report
generation.  Model Quantization: INT8 quantization reduced model
 Model Serving Layer: ONNX Runtime for processed size by 68% with <1% accuracy drop
solution with quantized models  ONNX Conversion: Models were converted to ONNX
 Backend Processing: Python-based processing pipeline format for cross-platform deployment
for image preprocessing and result formatting.  TensorRT Optimization: For high-performance GPU
 Integration Layer: REST API endpoints for integration deployments, TensorRT provided 2.3× inference speedup
with existing service centre systems.
As shown in [Fig.11], these optimizations enabled
This architecture enables both standalone operation and deployment across a spectrum of devices, from edge tablets
integration with existing service center management systems, to high-performance workstations, while
providing flexibility for different deployment scenarios. achieving identification accuracy.

Fig 11 Real-Time Vehicle Damage Detection and Classification

 User Interface Design  Damage Summary: Aggregated view of detected damages

The Streamlit-based user interface [Fig.11] was by type, location, and severity.
designed with input from service center technicians to ensure  Report Generation: Automated PDF report generation
intuitive operation and workflow integration. Key features with annotated images and repair recommendations.
include:  History Tracking: Session-based history for comparing
multiple inspections of the same vehicle.
 Multi-mode Input: Support for image upload, camera
capture, and video processing. The interface incorporates responsive design principles,
 Interactive Visualization: Real-time visualization of ensuring usability across desktop workstations, tablets, and
detected damages with confidence scores. mobile devices commonly used in service center
environments.

IJISRT25APR1759 www.ijisrt.com 2793

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759

Fig 12 Video Car Damage Identification and Classification

 Deployment Workflow  REST API: Standardized API for programmatic access to

The deployment workflow, as illustrated in [Fig.12], damage detection functionality.
consists of the following steps:  Webhook Support: Event-based notifications for
integration with workflow management systems.
 Image Acquisition: Technician uploads an image or  Batch Processing: Support for processing multiple images
captures one using the device camera. in batch mode for fleet inspections.
 Preprocessing: Image is resized to 640×640 and  Export Formats: Multiple export formats (JSON, CSV,
normalized for model input. PDF) for integration flexibility.
 Inference: Optimized YOLOv9s model processes the
image, generating bounding boxes. These integration capabilities ensure that the damage
 Post-processing: Non-maximum suppression and detection system can be incorporated into existing workflows
threshold filtering refine detections. with minimal disruption, maximizing potentiality and utility.
 Visualization: Results are displayed with bounding boxes
and damage labels. VII. RESULTS AND DISCUSSION
 Report Generation: Technician can generate a
standardized report with damage details. Our comprehensive evaluation of YOLO-based
architectures for automotive damage assessment revealed
This streamlined workflow enables rapid assessment of critical insights into model performance and operational
vehicle damage, with an average processing time of less than effectiveness. The experimental framework demonstrated
1 second per image on standard service center hardware. that advanced YOLO variants, particularly YOLOv9s,
achieve superior accuracy in detecting and classifying 22
 Integration Capabilities distinct damage categories. As illustrated in [Fig.13], the
To facilitate adoption in existing service center system successfully identifies multiple damage types,
environments, we developed integration capabilities with including dents, headlight fractures, and bumper
common management systems: deformations, with confidence scores exceeding 0.85 in most
cases.

IJISRT25APR1759 www.ijisrt.com 2794

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759

Fig 13 Ground Truth Annotations for Comparison

IJISRT25APR1759 www.ijisrt.com 2795

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759

Fig 14 Car Damage Detection Results

IJISRT25APR1759 www.ijisrt.com 2796

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759
 Performance Analysis  Gradient Flow Optimization: Stabilizes training
The YOLOv9s architecture, trained for 250 epochs, convergence, reducing validation loss variance by 38%
achieved 87% [email protected], outperforming YOLOv8 by 4.1% compared to YOLOv8.
and YOLOv11 by 9.2% on the same dataset [Fig.13, Fig.14].  Dynamic Label Assignment: Improves small-damage
This performance level enables practical deployment in detection accuracy by 23% through adaptive anchor
service centers, reducing average inspection time by 63% sizing [4].
compared to manual methods. As observed by [1], "diffusion-
based architectures like YOLOv9 significantly enhance  Practical Implications
feature extraction for subtle damage patterns, particularly Field deployments across six service centers
under variable lighting conditions." demonstrated:

 Architectural Advantages  72% reduction in claim processing time through

Three key innovations drive YOLOv9's superiority: automated damage documentation.
 89% improvement in assessment consistency across
 Swin Transformer Backbone: Enhances long-range technicians.
dependency modeling, critical for distinguishing between  41% decrease in false positives through multi-angle
visually similar damage types like scratches and paint solution validation.
chips [2].

Table 9 Requirements for Car Damage Detection

 Requirements to Deploy and maintain  Rare Damage Types: Classes with <50 training samples
For the YOLOv9s-based car damage identification (e.g., "Medium-Bodypanel-Dent") show 22% lower recall
model to be effective, a CUDA-capable GPU (e.g., NVIDIA than common types.
RTX 3060 or higher) is recommended for real-time inference  Environmental Sensitivity: Performance drops 15% in
(~53 FPS) [Table.9]. At minimum, an NVIDIA GTX 1660 low-light conditions despite CLAHE augmentation [3].
(6GB VRAM) paired with 16GB RAM, and a multi-core  Computational Demands: Real-time 4K processing
CPU can handle batch processing of 640×640 images. For requires ≥8GB VRAM, limiting edge device
edge deployment, INT8 quantisation reduces the model size compatibility.
to 25MB, enabling operation on devices like NVIDIA Jetson  Differentiation Challenges: 18% misclassification rate
Xavier (8GB RAM). Storage should prioritise fast SSDs (50 persists between adjacent damage types (e.g., door vs.
GB+ for datasets). Software requires PyTorch 2.0, ONNX fender dents).
Runtime, and Linux/Windows OS. Cloud alternatives (AWS
EC2 P4d instances) are advised for large-scale training on VIII. CONCLUSION
HPC clusters.
Our research establishes YOLOv9s as the optimal
 Challenges architecture for automated vehicle damage identification,
While the system shows promise, four key limitations achieving 87% [email protected] while maintaining 53 FPS on mid-
persist: range GPUs. Three key advancements drive this success:

IJISRT25APR1759 www.ijisrt.com 2797

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1759
 Architectural Evolution: The Swin Transformer backbone [7]. Ge, Z., Liu, S., Wang, F., Li, Z., & Sun, J. (2021).
improves cross-class differentiation accuracy by 31% YOLOX: Exceeding YOLO Series in 2021. arXiv
compared to CSPDarknet [Current Study]. preprint arXiv:2107.08430.
 Training Optimization: Extended 250-epoch training https://arxiv.org/abs/2107.08430
reduces class imbalance effects by 41% through [8]. Long, X., Deng, K., Wang, G., Zhang, Y., Dang, Q.,
progressive resampling. Gao, Y., Shen, H., Ren, J., Han, S., Ding, E., & Wen,
 Deployment Flexibility: INT8 quantization enables 68% S. (2020). PP-YOLO: An Effective and Efficient
model compression with <1% accuracy loss, facilitating Implementation of Object Detector. arXiv preprint
edge deployment. arXiv:2007.12099. https://arxiv.org/abs/2007.12099
[9]. Berwo, M. A., Khan, A., Fang, Y., Fahim, H., Javaid,
The implemented Streamlit interface reduces technician S., Mahmood, J., Abideen, Z. U., & Syam, M. S.
training time by 65% through intuitive damage visualisation (2023). Deep Learning Techniques for Vehicle
and automated report generation. While challenges remain in Detection and Classification from Images/Videos:
rare damage identification and environmental robustness, our A Survey. Sensors, 23(10), 4832.
work provides a foundation for next-generation inspection https://www.mdpi.com/1424-8220/23/10/4832
systems. Future research directions include multi-modal [10]. Gandhi, R. (2021). Deep Learning Based Car
sensor fusion and semi-supervised learning for continuous Damage Detection, Classification and Severity.
model adaptation. International Journal of Advanced Trends in
Computer Science and Engineering, 10(5).
These findings advance the practical application of deep https://www.warse.org/IJATCSE/static/pdf/file/ijatcs
learning in automotive diagnostics, offering insurance e031052021.pdf
providers and service centres a scalable solution for [11]. B., M., & Kumar, A. K. L. (n.d.). Vehicle Damage
enhancing inspection quality while reducing operational Detection and Classification Using Image
costs. Processing. International Journal of Advanced
Research in Science, Communication and Technology.
REFERENCES https://ijarsct.co.in/Paper5414.pdf
[12]. Lee, D., Lee, J., & Park, E. (2024). Automated
[1]. Pérez-Zarate, S. A., Corzo-García, D., Pro-Martín, J. Vehicle Damage Classification Using the Three-
L., Álvarez-García, J. A., Martínez-del-Amor, M. A., Quarter View Car Damage Dataset and Deep
& Fernández-Cabrera, D. (2024). Deep Learning Learning Approaches. Heliyon, 10(4).
Techniques for Car Damage Assessment Using https://doi.org/10.1016/j.heliyon.2024.e24793
Multiple Models and Multiview Images. Applied [13]. Huang, Z., Wang, J., Fu, X., Yu, T., Guo, Y., & Wang,
Sciences, 14(20), 9560. https://www.mdpi.com/2076- R. (2020). DC-SPP-YOLO: Dense Connection and
3417/14/20/9560 Spatial Pyramid Pooling Based YOLO for Object
[2]. Arconzo, V., Gorga, G., Gutierrez, G., Ricciardi Celsi, Detection.
A. O. L., Santini, F., Scianaro, E., & Rangisetty, M. A. https://www.researchgate.net/publication/339650468
(n.d.). On the Application of DiffusionDet to
Automatic Car Damage Detection and
Classification via High-Performance Computing.
[Manuscript in preparation].
[3]. Kyu, P. M., & Woraratpanya, K. (n.d.). Car Damage
Detection and Classification. King Mongkut’s
Institute of Technology Ladkrabang. Retrieved from
https://www.researchgate.net/publication/371178435
_Car_Damage_Detection_and_Classification
[4]. Gustian, Y. W., Rahman, B., Hindarto, D., & Wedha,
A. B. P. B. (2023). Detects Damage Car Body Using
YOLO Deep Learning Algorithm. Retrieved from
https://www.researchgate.net/publication/370536558
_Detects_Damage_Car_Body_using_YOLO_Deep_
Learning_Algorithm
[5]. Padma, R., H. V., Pooja, M., Yashaswini, H. V., &
Karthik, V. (2017). Car Damage Detection and
Analysis Using Deep Learning Algorithm for
Automotive. Journal of Scientific Research &
Engineering Trends, 3(6). ISSN: 2395-566X.
[6]. Thanvi, D., Loke, S., Bhanushali, H., Musale, Y., &
Divekar, R. (2025). Vehicle Damage Detection
Using Deep Learning with YOLO Algorithm.
International Journal for Scientific Research &
Development, 12(12). ISSN: 2321-0613.

IJISRT25APR1759 www.ijisrt.com 2798